\n| Home and Away by Mark Taylor – another interesting yet forgotten 80s console that also worked in the car. Notice the receipt, the price label in the style of Toys R Us, the console had everything and virtually sold very little…<\/span><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
<\/span><\/p>\nI suspect in time, AI will
\nbecome smart enough for us to be unable to distinguish whether something was
\ncreated using AI or not, and that could compound some of the issues we see
\ntoday when AI is used but not disclosed. \u00a0<\/span>Today, making the distinction between AI and
\nhuman art is possible with some visual training, but we shouldn\u2019t take it for
\ngranted that it will always be the case. AI is relatively still only at the
\nstart of its journey.<\/span><\/p>\nEven today, the images created
\nby AI are a huge leap from the days of Aaron, some of todays output is even
\npassable as artwork that wouldn\u2019t entirely look out of place in a gallery with
\nsome subjects looking almost indistinguishable from those drawn or painted by
\nhuman hands. They\u2019re certainly a long way away from the neural network images
\nthat became popular among mobile users in around 2016, an idea that created
\nabstract works based on photographs and images you had supplied and uploaded
\ninto the apps and very much a precursor to the technologies that we are seeing
\ntoday. <\/span><\/p>\nBut, there are some
\nlimitations with todays AI which to the casual viewer might not be immediately
\nvisible. Once you develop a slightly better understanding of AI\u2019s current image
\ncreating limitations you begin to notice the cracks and the similarities that
\nexist between AI works.<\/span><\/p>\nThere\u2019s the lack of detail and
\nthe distortions for a start, flaws that just wouldn\u2019t be present had the work
\nhave been created by a human, even an inexperienced human who has rarely
\ncreated art. The flaws look like computational errors which would, almost
\nironically, be very challenging for a human to create with the same effect. At
\nthe moment, so long as you have some basic knowledge of how AI images are
\nstructured and generated it\u2019s relatively straight forward to determine when
\nsomething is and isn\u2019t using AI. <\/span><\/p>\nIf you are aware of the
\nlimitations, it\u2019s possible even with a small amount of visual training to spot
\nimages created by AI, particularly images created with the more recent AI
\nengines such as Adobe\u2019s Firefly and DALL-E 2. If you have been exposed to human
\nproduced artworks over the course of an art career, the task of identifying AI
\nbecomes even easier. <\/span><\/p>\nAI images tend to also lack
\nthe emotion that will be more obvious in a work created by a human. AI can create
\nsome pretty soulless creations at times and we\u2019re maybe, even with the pace of
\ncurrent pace of change in the technology, still half a decade or more away from
\nAI that can replicate some of those personal nuances that humans bring to a
\nwork of art. Remember, AI is trained on data sets that have no idea of context
\nfrom a human perspective, that\u2019s also why we see so much bias occurring in some
\nof AIs output. What comes out of AI is only ever as good as what goes in, in
\nthis case, the data the AI is trained on.<\/span><\/p>\nWhen it comes to AI image
\ngeneration, today it feels like a by-product of everything else that AI can do.
\nAs it evolves, for artists, it will eventually serve one of two purposes, it
\nwill either help enormously with the creative process or it will become the
\nforce that displaces an entire industry, which road it takes will be largely
\ndetermined by how artists begin to embrace it, or move away from it, and to
\nsome extent, the public\u2019s acceptance of AIs role in creating art.<\/span><\/p>\nOne of the big problems with
\nAI image generation today is that it relies on what I\u2019ve come to term as the
\nbillboard effect. When you look at an image created by a human, be it a
\nphotograph or some other image, what you look at tends to be in high
\nresolution. Mostly, printed images are created by thousands upon thousands of
\nsmall dots printed by an inkjet printer or represented by pixels on a screen. In
\nshort, most images that you will be familiar with, have detail that is clear
\neven when looking at the images for a prolonged period of time.<\/span><\/p>\nBillboards can be printed
\nusing very low resolutions, these images are created with far fewer dots and
\nnowhere even close to having the same level of detail that would be found in a
\nfine art print. The billboard images work because we\u2019re viewing them, often only
\nbriefly, from a distance away. <\/span><\/p>\nOnce we get around 650 feet
\naway from an image, our eyes can only resolve around one pixel per inch. While
\n300 dots per inch seems to be the golden rule for print resolution, in truth,
\nit\u2019s not a super helpful number that can be applied to everything and it also
\nmakes prints more expensive to produce. But, it is an accepted standard that
\nreproduces the detail needed for the best fine art prints, it\u2019s also
\nunforgiving for artists because at that resolution, any discrepancies in the
\nwork are easier to spot.<\/span><\/p>\nA billboard is usually printed
\nat 15dpi, the advertising on the side of a bus is usually no more than
\n72-100dpi, and many fine art prints are usually printed at 240dpi, although I
\ndo tend to stick to the regular 300dpi because I put so much effort into
\ncreating intricate, even small details into my retro works and some of the
\nprints I offer are sold at a size that would be large enough to notice the
\ndifference up close. \u00a0<\/span>AI models found
\nonline are usually creating work at a screen resolution of somewhere in the
\nregion of 72 dots per inch. In short, not really good enough to hang on the
\nwall.<\/span><\/p>\nSo, when you look at low
\nresolution images on a billboard from a distance you are doing the heavy
\nlifting by resolving the missing detail. You generally only ever view a
\nbillboard for a short period of time and you are usually standing too far away
\nto be able to view any detail even if it was present because of the shortcomings
\nin not being able to process more than about one pixel per inch from a distance.
\n<\/span><\/p>\nWith AI images, it\u2019s kind of
\nthe same thing. It\u2019s an approximation of an image, but get up close and study
\nit for a few seconds longer and the cracks begin to show. We might stumble
\nacross AI images in the news or online, both of which would be providing us
\nwith images that have much lower resolutions than would be expected from a fine
\nart print, but we will usually move on much more quickly than if we were
\nviewing a high resolution artwork that has been professionally printed and hung
\non the wall at home or in a gallery where we\u2019re also closer to the work. AI
\nimage generation is more smoke and mirrors than we might initially think it is.<\/span><\/p>\nAI simply doesn\u2019t do detail
\nvery well when it comes to images. It also struggles with other things too, so
\nwe will often find misquoted information or information that makes no sense and
\nit hallucinates more than a hippy on a mushroom trip at times. When AI is
\nfocussed on a particular task and is less generalised, it tends to perform way better,
\nhence medical research which is predicated on learning from those very specific
\nsmall data models, which also make it generally much better and faster than a
\nhuman.<\/span><\/p>\nWhen AI becomes part of a
\nconsumer facing system that tries to be everything to everyone, it struggles. In
\npart, unlike humans who are usually inspired by reality, AI is inspired by
\nwhatever dataset it has been trained on and it literally doesn\u2019t know anything else.
\nAt best, this means that it can only ever create an approximation or
\nreproduction because it has no real context or reality to base the output on. Until
\nthe issues of context and emotion are included in the data sets, I think AI art
\nwill continue to struggle because there will be missing data and missing detail
\nthat humans understand because they have life experience to baseline it on.<\/span><\/p>\nIf you look at AI generated
\nhands, they tend to be deformed for one reason alone, the datasets containing
\nthe images used to train the AI will be more likely to have featured faces, and
\nfaces tend to be much more prominent than hands when you look through any photo
\ncollection.<\/span><\/p>\nThat said, if we look at how
\nmuch hand generation has improved even since the beginning of this year, we are
\nnow getting to a point where even hands are becoming more identifiable as
\nhands. It\u2019s not that AI is getting better, it\u2019s that the data sets are getting
\nbigger and more varied. There are still plenty of tells though, the occasional
\nextra thumb, hands appearing in an anatomically incorrect place, these are the tells
\nthat AI is being used.<\/span><\/p>\nUnnatural textures, either too
\nsmooth or perfect or that don\u2019t look quite right or are shown in the wrong
\ncontext can also be an obvious tell, as can plastic-like skin and unnatural
\nskin tones, or the ultimate giveaway this month, retro-futuristic colour
\npalettes. When recreating materials, often the materials will look overly
\nuniform, there may be no creases or shadows. If there are obvious brush
\nstrokes, look for replication of the same brush stroke within the image, this
\nwould imply that a digital brush algorithm is being used.<\/span><\/p>\nThere will be other inconsistencies
\nor objects that appear to be out of place, any light refraction included in an
\noriginal photograph used to train the AI is likely to still be present in the output
\nand it is often represented by weird black dots and lines. Lighting is another
\ngiveaway, a human artist will instinctively know that reflections are generally
\nuniform and light is cast often from a single point, especially in landscapes
\nwhen either the sun or moon is the light source, AI generally doesn\u2019t
\nunderstand how to render light sources that well so it casts misplaced shadows
\nand reflections.<\/span><\/p>\nIf you zoom into the image,
\nyou might notice unnatural poses and expressions, stark colour changes that
\nmake no sense, artifacts left behind from the original source, pixelation in
\nsome or all areas of the image, or an halo effect that looks like it should be
\nfixed with an application of gaussian blur in Photoshop. You will especially
\nsee this in fake images of anything usually in flight. UFO photographs can
\noften be debunked within seconds when AI has been used.<\/span><\/p>\nAnother giveaway is that the composition
\nof the image tends to be disjointed, or the image may be unnecessarily cropped
\nin a way that makes little to no sense. Artists are usually taught composition or
\nthey will pick it up as they gain experience, AI has certainly got a way to go
\nbefore composition becomes natural, partly because it has no reference or
\ncontext other than the training images it has been trained on. Size and scale
\nare difficult to comprehend in a photograph unless you have some other context
\nor other point of reference.<\/span><\/p>\nRepetition and unnatural
\npatterns are another area where AI struggles. Misaligned repetition of a
\npattern is often the first clue, but also look out for patterns that would
\nfirst appear to fit together if they were to be realigned, but would then still
\nnot match perfectly. AI, is still incapable of creating perfection. <\/span><\/p>\nYou might also want to conduct
\na reverse image search of the image in its entirety and sections of it. \u00a0<\/span>The reverse image algorithms will be typically
\nlooking at the same reference photos online that had been used to train the AI,
\nso it\u2019s relatively easy for them to pick out elements of an image that may be
\nfrom another artists work. <\/span><\/p>\nIt\u2019s also worth looking for
\nthe artists signature, if it\u2019s present and in the usual place (most artists
\nwill sign their work in the same way, usually in the same area of their work), you
\nshould look for smoothing of the area where the image looks like it has either
\nbeen healed or cloned from another section of the image. In instances where the
\nsignature or watermark has been removed smooth areas tend to mean that the
\nimage has been manipulated and sometimes the area looks as if it has had a
\nsection of the surrounding area pasted over the signature. <\/span><\/p>\n\n\n\n![\"retro](\"https:\/\/blogger.googleusercontent.com\/img\/b\/R29vZ2xl\/AVvXsEhsEccmyL7SmBimvQxDalHA74IpC_K_k5YnEbM4RT6pEiey7aYMeV6EGTtL-45KweEjbRo246p3BWYSHoZXERXcU3mQxyfbiOyFIF-mSw1q7iH0KiHkjxTHU_JWUJvnc3Lc5WhoGCEFIj7vripFDLjrWS-vLSueWbimDSwgELGSN-UaMa8kKf5GqtUBkXc\/w640-h640\/image%206.JPG\" "\\"Forgotten") <\/a><\/td>\n<\/tr>\n\n| Forgotten From Taiwan by Mark Taylor – this sold better than others, today it is collectible, again, plenty of Easter Eggs to find in my latest works…<\/span><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n
<\/span><\/p>\nYou might be wondering how
\neasy all of this is, but on a recent social media scroll I managed to pick out
\na dozen or so images that had definitively been created with the use of AI, and
\nnone had been disclosed as being produced using artificial intelligence. Just
\nto be clear, there was no need for me to perform any kind of forensic
\nshenaniganry to categorically prove each image was AI, it was evident visually
\nbut backed up with a screen shot and reverse image search. The whole process,
\nwell it took me less than 10-minutes.<\/span><\/p>\nWhat gave the game away for
\nthese images were that two were juxtapositions of another artists work which I
\nwas familiar with, and both were essentially the same image. Another used a
\ncolour pallet that would only make sense with an entirely different subject and
\nthere was no smooth gradation of colour, it was stark and pixelated in a way
\nthat wouldn\u2019t indicate that the image had been badly resized. Out of the ten
\nimages, a number had been uploaded to a print on demand service and were
\navailable for sale, again with no disclosure of the process of using AI in the
\nmarketing description. If you are familiar with the capabilities of any of the
\nmajor digital art applications such as Photoshop, Corel, or Serif, that alone
\nwill give you a good foundation on which you can become a master sleuth in this
\nfield.<\/span><\/p>\nMy experience which spans back
\nto day one of Photoshop and before that, Delux Paint on the Commodore Amiga
\ncomputer and even earlier 8-bit micro\u2019s in the 80s and their rudimentary
\nimaging programs, and having used almost every digital art application since,
\nthis long-term experience \u00a0<\/span>means that I
\nam probably in the unique position of being an official anorak in this
\ndepartment. I can tell the difference between photoshop being used and Corel,
\nbut with only minimal visual training I really do believe that almost anyone
\ncan currently master this dark art of distinguishing an AI image from the real
\nthing, and I would certainly encourage art buyers today to be extra cautious
\nwhen buying work they believe to have been created solely by the artist.<\/span><\/p>\nThat brings me nicely to
\nanother point. Most savvy art buyers will already be looking out for these
\nvisual cues and it\u2019s not just art buyers who buy the most expensive works. As
\nAI becomes more widely accessible, that means that those people who currently
\nbuy work have the same level of access as anyone else to the same AI tools.
\nSome of the more recent conversations I have had with my own buyers mostly
\nconfirm that they\u2019re quickly becoming adept at spotting when something is amiss
\nwith a piece of work, and more of them are rightfully asking the right
\nquestions around whether AI had been used in the creation. <\/span><\/p>\nWhen AI art is generated by a
\nperson who has minimal experience with AI prompts, I suspect that some of this
\nmissing detail is also the result of the artist having a lack of understanding around
\nhow AI algorithms are programmed. The models AI image generators are being
\ntrained on are getting better, but the results we\u2019re still seeing from
\ninexperienced AI artists aren\u2019t quite reaching their true potential just yet. <\/span><\/p>\nTo get better results you need
\nto start out with very specific instructions and additional modifiers need to
\nbe progressively included to change the conditions on which the final image is
\ngenerated. Having an understanding of composition and artistic styles would be
\nuseful here, without that background knowledge, AI derived art will always look
\na bit soulless and often generic.<\/span><\/p>\nIt\u2019s having that extra level
\nof art and design knowledge that would make an AI image more believable, but
\nmany who are casually creating art on these platforms will take whatever comes
\nout. The instructions we type into AI are called prompts, but prompt modifiers
\ncan be used to refine the results. If the modifiers are grounded in the user
\nhaving a knowledge of art history, the modifiers can be made much better and
\nthe results will be much more believable.<\/span><\/p>\nYour subject is what will be
\nincluded in the image, it becomes the main focal point, so a modifier would not
\nonly include the subject but would also include for example, a specific
\nbackground, a particular location, or an action. Modifiers give the subject
\nadditional context because without the additional context provided by a human,
\nAI is generally very prescriptive. Ask it for an image of a frog and you will
\nget an image of a frog, but ask it for an image of a frog sitting on a tree
\nbranch in the jungle surrounded by exotic flowers and you then have much more
\ncontext which the AI can work with to level up what could otherwise be a rather
\nbland image of a lonely frog.<\/span><\/p>\nYou can further refine the
\nprompt modifiers to include things like actions, so the frog could also be
\nreading a book. If you think back to the lesson on verbs during your days at
\nschool, words such as: fall, run, jump, push, pull, play, sit, stand, can all
\nbe used to give even more context to the prompt, each time adding a further
\nelement of detail on which the AI can do its thing.<\/span><\/p>\nBeyond actions, you could also
\napply additional modifiers such as giving it the context to produce the work in
\na specific artistic style, this could be photographic or abstract, equally it
\ncould be panoramic or it could be in the style of a particular artist. If the
\nmodel has been trained with images reflecting those prompts, the output again becomes
\nmore believable. <\/span><\/p>\nMacrophotography is something
\nthat can be recreated well with AI, there\u2019s less detail in the background to
\nget wrong, but what you are more likely to see from an inexperienced AI user
\nwith little to no artistic knowledge is usually the same image that anyone
\ncould generate with a very simple prompt. That\u2019s another way of figuring out
\nwhat\u2019s AI and what\u2019s more likely to be original, try and recreate the same
\nimage using a simple description, chances are, the output you will see will be
\nthe same or a very similar image.<\/span><\/p>\nPrompt modifiers for image
\ngeneration can even extend to materials and mediums, lighting, colour palette,
\nperspective, mood, or even an era or historic period in time. <\/span><\/p>\n
| |
|
![\"Artistic](\"https:\/\/blogger.googleusercontent.com\/img\/b\/R29vZ2xl\/AVvXsEhYiZMJ5YxBA6jcvXZ1nKGdhRzMXWn08SAhg6KCVKzlIN6qVwe3KQqVBqPXdnxgTWDYbSQSQaNu9btuBfbRpJQ2Ui2_zU_pxP6vxY27kVjt75YaMZSV-7hbB2FA_QIfaEGahYbTV_TE9wZ93JkwR9J77ezc90owzbgCznItuK28Nr1jyzsgOt-WsbWk13M\/s2356\/artistic%20integrity%20cover.PNG\")
<\/a><\/td>\n<\/tr>\n